Error-Tolerant Interactive 2D-to-3D Conversion Using Local Consistency

Abstract

Existing interactive 2D-to-3D methods assume that user input is perfectly accurate. However, it is hard to get perfectly accurate user input and even small errors will degrade the conversion quality. To address this problem, user scribble confidence is proposed to remove input errors by using local consistency between labeled pixels and their neighbors. First, we count the number of neighbors which have similar and different color values for each labeled pixels respectively. The ratio between these two numbers at each labeled pixel is regarded as its scribble confidence. Second, 2D-to-3D conversion is formulated as a confident optimization problem by introducing a confident weighting data cost term, the local and K-nearest depth consistent regularization terms. The proposed method is compared with the state-of-the-art methods on several representative images. The experimental results demonstrate that our method can tolerate input errors and PSNR is improved by more than 1 dB compared with existing algorithms when inaccurate scribbles are present.